triage-ml-task

Triage ML Task

This is the session owner. Stage skills do the work; you only route and ask. Do not execute another skill's methodology.

Human-facing prose

Details: setup-workspace references/human_facing_prose.md. Questions and replies describe the work (explore the data, build a model) — not skill ids, G-* names, or the wrapper CLI. Run python -m skore_skills … yourself; do not quote it.

Every question here carries its own context: 2–4 lines on what the answer authorizes, the workspace facts it rests on — echoed inline from status (scaffold, data_analysis, loop_stage) — and what each option leads to. A file link is an addition, never the context.

Procedure

1. Run python -m skore_skills status. Read skills, data_analysis,

loop_stage, setup.pending, and the filesystem snapshot. If .skore is missing, that is expected. Do not treat a missing file as an empty project when src/ or journal/ exist.

1. Certain request — load that skill, after the setup gate

below. Tell the user the work you are starting, not the catalog id. Do not list the catalog. status.skills is a per-id dict. Load the mapped skill only if status.skills.<id> is true; else one-line skip and do not invent that skill's steps:

Setup gate for a lifecycle skill. Before loading explore-ml-data, frame-ml-problem, model-ml-pipeline, build-ml-pipeline, evaluate-ml-pipeline, smoke-test-ml-pipeline, or audit-ml-pipeline, read status.setup.pending. If it is non-empty and status.skills.setup-ml-project is true, load setup-ml-project and stop. Tell the user the requested work waits on those pieces. Do not show the entry menu. After setup returns, load the certain lifecycle skill. Do not load setup again on this turn. If setup.env or setup.workspace is declined, stop in one line. A declined git or editable does not block. If setup-ml-project is not installed, name the pending pieces in one line and stop. Do not invent git init, scaffold, or env init. Uncertain sessions skip this gate.

An explicit sync, generic export, or changed modeling constraint still uses those rows. Do not send them through review-ml-choices.

A git init, first commit, or ignore request loads setup-git and stops. Do not write git init, a .gitignore, or a commit plan. With no shell, the whole answer is that setup-git is loaded. Do not describe the commands that skill will run.

"What should we try next?" while loop_stage is backlog and manage-ml-backlog is installed loads that skill and stops. Do not ask the user to choose explore, build, review, or export.

Certain EDA: run python -m skore_skills status. If status.setup.pending is non-empty, load setup-ml-project (setup gate above) and stop. Otherwise load explore-ml-data and stop. Do not inventory data/, list missingness or distributions, or start EDA methodology.

Certain generic export: run python -m skore_skills status, then load export-ml-project. Do not list notebook / --html / site as sibling options.

Modeling while `status.data_analysis` is `missing`: if the certain skill is model-ml-pipeline (or the user asked to build the first experiment) and status.skills.explore-ml-data is true, do not load modeling yet. AskUserQuestion two options: run exploratory data analysis first (default) vs proceed to modeling with user-supplied facts. Do not invent dataset facts here. If data_analysis is present or skipped, load model-ml-pipeline with no extra gate (still only if that id is true).

1. Uncertain (open session, “what can you do”, finished stage,

mixed intent) — AskUserQuestion with the installed entry work only. One pick, then load the mapped skill.

Offer a label only if status.skills.<id> is true. User-visible labels (no ids):

• Set up the project → setup-ml-project
• Explore the data → explore-ml-data
• Build a model → model-ml-pipeline
• Review the last experiment → review-ml-experiment
• Record / decide what next → manage-ml-backlog
• Export → export-ml-project
• Sync reports → sync-ml-reports
• Review choices → review-ml-choices

If none of those ids are true, say so in one line; do not invent a menu.

Do not put evaluate or audit on this board (certain requests still load evaluate-ml-pipeline / audit-ml-pipeline). Do not put shaping an idea or literature search on this board (certain requests and manage-ml-backlog still load shape-user-idea / search-ml-literature).

If status.data_analysis is missing and status.skills.explore-ml-data is true, recommend exploring the data first. Do not auto-load it. When data_analysis is present or skipped and loop_stage is backlog, recommend recording the run / deciding what next when manage-ml-backlog is true. Do not auto-load it.

Do not put internals on this board (build-ml-pipeline, smoke-test-ml-pipeline except as a certain debug load, frame-ml-problem, choose-python-library, research-ml-practice, plot-ml-figure, stack refs).

Stop conditions

• Do not design experiments, write pipelines, or run exploratory

data analysis yourself. Certain EDA with setup.pending empty loads explore-ml-data only — no data inventory and no EDA checklist. Pending setup loads setup-ml-project first.

• Do not load every skill.
• Do not invent workspace facts when status is unavailable.
• Do not treat a missing .skore as an empty project when src/

or journal/ exist.

• Do not invent a missing skill's steps from memory.
• Do not put evaluate-ml-pipeline or audit-ml-pipeline on the

uncertain entry board (certain requests still load them).

• A missing review, backlog, user-idea, or literature skill is

a one-line skip. Do not invent that skill's procedure.

• Do not put shape-user-idea or search-ml-literature on the

uncertain entry board.

• Do not put frame-ml-problem on the uncertain entry board

(a certain metric, split, baseline, or changed-constraint request still loads it).

• Certain generic export: run python -m skore_skills status,

load export-ml-project only — no sibling-skill menu.

End of every other skill's turn returns here when this skill is installed.